Generative recommendation (GR) reformulates item retrieval as autoregressive generation of discrete semantic identifiers (SIDs), offering a unified paradigm for personalized recommendation. GR4AD, as an industrial generative recommendation framework, has greatly improved advertising revenue while enabling efficient real-time serving in Kuaishou.However, continued deployment has exposed three limitations: deterministic SID assignment remains sensitive to semantic boundaries; conventional backbone scaling provides diminishing revenue gains under strict serving constraints; and sequence-level value optimization provides limited credit assignment to intermediate SID decisions, causing likelihood-based beam search to prematurely prune potentially high-value candidates. We present GR4AD-X, which addresses these limitations through reliable SID construction, serving-efficient capacity scaling, and unified value alignment. Specifically, GR4AD-X improves SID reliability through Multi-Path assignment. It expands model capacity along multiple complementary dimensions rather than relying solely on wider or deeper backbones, achieving a better effectiveness-efficiency trade-off under strict serving constraints. Finally, Unified Value-Guided Generation and Discrimination (UniVGD) aligns generation with commercial value through prefix-level reward optimization, value-guided beam pruning, and discriminative reranking within a unified model. Large-scale online A/B tests on the Kuaishou advertising platform demonstrate that GR4AD-X achieves a cumulative revenue improvement of 3.07% over the deployed GR4AD baseline while satisfying strict latency and throughput constraints.
Figures & tables
Figure 1. Production trade-off of naive backbone scaling. All variants use Multi-Path SID. Serving QPS and revenue improvements are reported relative to the 0.03 B variant. Larger backbones substantially reduce serving throughput, while revenue gains exhibit diminishing returns.
Figure 2. Overview of GR4AD-X. (a) Multi-Path SID retains multiple plausible first-level assignments for boundary ads. (b) PV-BeamSearch expands to 2k candidates and fuses generation and prefix-value ranks via RRF. (c) Revenue rewards propagate bottom-up through the SID prefix tree for PV-RSPO. (d) The backbone uses a Shared Decoder with diagonal Hyper-Connections, a Step Decoder with PCGM, and per-level prefix-value heads for value alignment and discriminative reranking.
Model Settings
Δ Revenue vs. GR4AD
Reliable SID Construction
+ Multi-Path SID
+0.58%
Efficient Model Scaling
+ Input & Encoder Scaling
+0.96%
+ Vocabulary Scaling (PCGM)
+1.43%
+ Stream Scaling (Diag. HC)
+1.72%
Table 1. Sequential online A/B results for GR4AD-X. Components are added cumulatively, and each row reports the cumulative revenue improvement over the deployed GR4AD baseline.
Figure 3. Geometric motivation and first-level target retention of Multi-Path SID. (a) Distribution of the distance ratio r=d1/d2 over 7.556M advertisement embeddings with 16,384 first-level centroids. (b) First-level hit rate of Single-Path and Multi-Path SID at K∈{1,16,64} . Labels above each pair indicate relative improvements over Single-Path SID.
Variant
Δ Revenue
Parameters
QPS
Multi-Path SID baseline
–
0.16 B
100%
+ Input scaling
+0.20%
0.16 B
100%
+ Encoder scaling
+0.38%
0.20 B
90%
+ Vocabulary scaling
+0.85%
0.46 B
95%
+ Stream scaling
+1.13%
0.49 B
90%
+ Decision-depth scaling
+1.41%
0.49 B
85%
Table 2. Sequential online A/B results of serving-efficient capacity scaling. Each proposed dimension is added on top of the preceding variant. Revenue improvement and QPS are normalized to the 0.16 B Multi-Path SID baseline. The naive scaling result is derived from the same Multi-Path SID-enabled experiments in Figure 1 , re-normalized to the 0.16 B baseline.
Base Model
+ UniVGD
Lvl-1
Lvl-1&2
Full
Lvl-1
Lvl-1&2
Full
HR@10
0.017
0.019
0.021
0.145
0.287
0.298
HR@50
0.071
0.075
0.078
0.194
0.329
0.323
HR@100
0.216
0.224
0.221
0.306
0.441
0.427
Table 3. HitRate@ K against eCPM at three SID scopes. Base uses generation probability; UniVGD uses probability at Lvl-1 and the RRF beam-selection score thereafter. Full excludes final reranking.
Generative recommendation has recently attracted widespread attention in industry due to its potential for scaling and stronger model capacity. However, deploying real-time generative recommendation in large-scale advertising requires designs beyond large-language-model (LLM)-style training and serving recipes. We present a production-oriented generative recommender co-designed across architecture, learning, and serving, named GR4AD (Generative Recommendation for ADdvertising). As for tokenization, GR4AD proposes UA-SID (Unified Advertisement Semantic ID) to capture complicated business information. Furthermore, GR4AD introduces LazyAR, a lazy autoregressive decoder that relaxes layer-wise dependencies for short, multi-candidate generation, preserving effectiveness while reducing inference cost, which facilitates scaling under fixed serving budgets. To align optimization with business value, GR4AD employs VSL (Value-Aware Supervised Learning) and proposes RSPO (Ranking-Guided Softmax Preference Optimization), a ranking-aware, list-wise reinforcement learning algorithm that optimizes value-based rewards under list-level metrics for continual online updates. For online inference, we further propose dynamic beam serving, which adapts beam width across generation levels and online load to control compute. Large-scale online A/B tests show up to 4.2% ad revenue improvement over an existing DLRM-based stack, with consistent gains from both model scaling and inference-time scaling. GR4AD has been fully deployed in Kuaishou advertising system with over 400 million users and achieves high-throughput real-time serving.
Generative Recommender Systems using semantic ids, such as TIGER (Rajput et al., 2023), have emerged as a widely adopted competitive paradigm in sequential recommendation. However, existing architectures are designed solely for semantic retrieval and do not address concerns such as monetization via ad revenue and incorporation of bids for commercial retrieval. We propose GEM-Rec, a unified framework that integrates commercial relevance and monetization objectives directly into the generative sequence. We introduce control tokens to decouple the decision of whether to show an ad from which item to show. This allows the model to learn valid placement patterns directly from interaction logs, which inherently reflect past successful ad placements. Complementing this, we devise a Bid-Aware Decoding mechanism that handles real-time pricing, injecting bids directly into the inference process to steer the generation toward high-value items. We prove that this approach guarantees allocation monotonicity, ensuring that higher bids weakly increase an ad's likelihood of being shown without requiring model retraining. Experiments demonstrate that GEM-Rec allows platforms to dynamically optimize for semantic relevance and platform revenue.
Yanchen Jiang, Zhe Feng, Christopher P. Mah +2
Harvard University · Work done during an internship at Google Research · Google Research
Generative Retrieval (GR) offers a promising paradigm for recommendation through next-token prediction (NTP). However, scaling it to large-scale industrial systems introduces three challenges: (i) within a single request, the identical model inputs may produce inconsistent outputs due to the pagination request mechanism; (ii) the prohibitive cost of encoding long user behavior sequences with multi-token item representations based on semantic IDs, and (iii) aligning the generative policy with nuanced user preference signals. We present GenRec, a preference-oriented generative framework deployed on the JD App that addresses above challenges within a single decoder-only architecture. For training objective, we propose Page-wise NTP task, which supervises over an entire interaction page rather than each interacted item individually, providing denser gradient signal and resolving the one-to-many ambiguity of point-wise training. On the prefilling side, an asymmetric linear Token Merger compresses multi-token Semantic IDs in the prompt while preserving full-resolution decoding, reducing input length by ~2X with negligible accuracy loss. To further align outputs with user satisfaction, we introduce GRPO-SR, a reinforcement learning method that pairs Group Relative Policy Optimization with NLL regularization for training stability, and employs Hybrid Rewards combining a dense reward model with a relevance gate to mitigate reward hacking. In month-long online A/B tests serving production traffic, GenRec achieves 9.5% improvement in click count and 8.7% in transaction count over the existing pipeline.
Yanyan Zou, Junbo Qi, Lunsong Huang +7
JD.com Beijing, China · Waseda University Tokyo, Japan